Question Answering over Freebase with Multi-Column Convolutional Neural Networks

Proceedings of the 53rd Annual Meeting of the Association for Computational Linguistics and the 7th International Joint Conference on Natural Language Processing, pages 260–269, Beijing, China, July 26-31, 2015. c⃝2015 Association for Computational Linguistics Question Answering over Freebase with Multi-Column Convolutional Neural Networks Li Dong†∗ Furu Wei‡ Ming Zhou‡ Ke Xu† †SKLSDE Lab, Beihang University, Beijing, China ‡Microsoft Research, Beijing, China donglixp@gmail.com {fuwei,mingzhou}@microsoft.com kexu@nlsde.buaa.edu.cn Abstract Answering natural language questions over a knowledge base is an important and challenging task. Most of existing sys- tems typically rely on hand-crafted fea- tures and rules to conduct question under- standing and/or answer ranking. In this pa- per, we introduce multi-column convolu- tional neural networks (MCCNNs) to un- derstand questions from three different as- pects (namely, answer path, answer con- text, and answer type) and learn their dis- tributed representations. Meanwhile, we jointly learn low-dimensional embeddings of entities and relations in the knowledge base. Question-answer pairs are used to train the model to rank candidate answers. We also leverage question paraphrases to train the column networks in a multi-task learning manner. We use FREEBASE as the knowledge base and conduct exten- sive experiments on the WEBQUESTIONS dataset. Experimental results show that our method achieves better or comparable performance compared with baseline sys- tems. In addition, we develop a method to compute the salience scores of question words in different column networks. The results help us intuitively understand what MCCNNs learn. 1 Introduction Automatic question answering systems return the direct and exact answers to natural language ques- tions. In recent years, the development of large- scale knowledge bases, such as FREEBASE (Bol- lacker et al., 2008), provides a rich resource to answer open-domain questions. However, how ∗Contribution during internship at Microsoft Research. to understand questions and bridge the gap be- tween natural languages and structured semantics of knowledge bases is still very challenging. Up to now, there are two mainstream methods for this task. The first one is based on seman- tic parsing (Berant et al., 2013; Berant and Liang, 2014) and the other relies on information extrac- tion over the structured knowledge base (Yao and Van Durme, 2014; Bordes et al., 2014a; Bordes et al., 2014b). The semantic parsers learn to un- derstand natural language questions by converting them into logical forms. Then, the parse results are used to generate structured queries to search knowledge bases and obtain the answers. Re- cent works mainly focus on using question-answer pairs, instead of annotated logical forms of ques- tions, as weak training signals (Liang et al., 2011; Krishnamurthy and Mitchell, 2012) to reduce an- notation costs. However, some of them still as- sume a fixed and pre-defined set of lexical trig- gers which limit their domains and scalability ca- pability. In addition, they need to manually de- sign features for semantic parsers. The second approach uses information extraction techniques for open question answering. These methods re- trieve a set of candidate answers from the knowl- edge base, and the extract features for the question and these …

Content truncated to 500 words